Participant grouping method and device for federated learning of packets and electronic equipment

By analyzing the data distribution characteristics and calculating the weights of the participants in grouped federated learning, feature fingerprints are generated, resulting in more reasonable grouping results. This solves the problem of inconsistent grouping results and improves the accuracy and efficiency of model training.

CN122432706APending Publication Date: 2026-07-21NINGXIA CREDIT INFORMATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NINGXIA CREDIT INFORMATION CO LTD
Filing Date
2026-04-23
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

In existing technologies, the grouping results of multiple participants in grouped federated learning do not match the data distribution characteristics of the actual data held by each participant, resulting in poor model generalization ability, slow convergence speed, and even training divergence.

Method used

By analyzing the data distribution characteristics of local data, feature fingerprints are generated, and the affiliation weight calculation model is used to calculate the posterior probability of each client belonging to each group. Grouping is performed based on the posterior probability, allowing a client to participate in the training of multiple global sub-models simultaneously. A weighted soft routing mechanism is used for grouping.

Benefits of technology

It improves the rationality of grouping results, solves the problem that participants in the classification boundary region are difficult to be accurately grouped, and enhances the accuracy and efficiency of model training.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122432706A_ABST
    Figure CN122432706A_ABST
Patent Text Reader

Abstract

The application discloses a participant grouping method and device for packet federal learning and electronic equipment, relates to the technical field of machine learning and the technical field of multi-party secure computation, wherein a client of each participant in a plurality of participants obtains a feature fingerprint of local data by performing data distribution feature analysis on the local data, the feature fingerprint includes a feature vector representing the data distribution feature, and is sent to a server; the server calculates a posterior probability of each client belonging to each group based on the feature fingerprints from the plurality of clients by using a belonging degree weight calculation model, takes the posterior probability as a belonging degree weight, and determines each client included in each group based on a size relationship between the belonging degree weight and a preset weight threshold. By using the scheme, the problem that participants in a classification boundary region are difficult to be accurately grouped and modeled in model training is solved, that is, the rationality of the grouping result is further improved.
Need to check novelty before this filing date? Find Prior Art